Agent Architectures in Practice: Patterns, Platforms, and Use Cases
Choose the simplest approach that handles the task, then test its failure paths.
For the people deciding what to build, and the people making it work. Practical notes, honest trade-offs and small projects from Jhony Vidal.
Find your next readYour support tool has the order, the history and the policy. Why make someone explain it all to an AI assistant again?
Give a familiar task a useful action, with a result the person can check.
“I’ve checked your order. We’re waiting for a confirmed delivery date…”
Three ways into the notes.
What should the feature do, and what should a person stay in control of?
Start with assistant memoryFind the sources, inspect the mistakes and learn whether a change helped.
Start with evaluating answersChoose what deserves your attention as the tools and the work change.
Try a two-week learning loop13 notes to explore
Choose the simplest approach that handles the task, then test its failure paths.
Review tasks within a job and test what helps before making assumptions about whole roles.
Compare useful completed work, including mistakes, review time and retries.
Test finding evidence separately from writing an answer supported by that evidence.
Choose one recurring problem, practise recall and return to it after a gap.
Turn a repeated task into a useful action, with a reply the person can check and change.
A job runs, the connection drops, and someone tries again. Does the total change? Try two fictional events and see what should happen.
Explore the projectevt-1 £12.00evt-2 £10.49Ready. Process the two events to begin.
I’m a software engineer working with AI, data and automation. I write down the decisions, examples and mistakes that are worth returning to. The goal is to leave you with something you can understand, question or try in your own work.
